METHOD AND SYSTEM FOR THE SPOTTING OF ARBITRARY WORDS IN HANDWRITTEN DOCUMENTS

• US 20150269431A1
• Filed: 11/19/2013
• Published: 09/24/2015
• Est. Priority Date: 11/19/2012
• Status: Active Grant

• Current Assignee: IMDS Group Inc.
• Original Assignee: IMD America Inc.
• Inventors: Haji, Mehdi, Ponson, Dominique
• Granted Patent: US 9,740,925 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06V 30/2276 with probabilistic networks, e.g. hidden Markov models
• CPC Class Codes: G06V 30/36 Matching; Classification
• CPC Class Codes: G06V 40/30 Writer recognition; Reading and verifying signatures

Abstract

A method and system for the spotting of keywords in a handwritten document, the method comprising the steps of inputting an image of the handwritten document, performing word segmentation on the image to obtain segmented words, performing word matching, and outputting the spotted keywords. The word matching itself consisting in the sub-steps of performing character segmentation on the segmented words, performing character recognition on the segmented characters, performing distance computations on the recognized characters using a Generalized Hidden Markov Model with ergodic topology to identify words based on character models and performing non-keyword rejection using a classifier based on a combination of Gaussian Mixture Models, Hidden Markov Models and Support Vector Machines.

Claims

1. A method for the spotting of keywords in a handwritten document, comprising the steps of:

inputting an image of the handwritten document;

performing word segmentation on the image to obtain segmented words;

performing word matching, consisting in the sub-steps of;

performing character segmentation on the segmented words;

performing character recognition on the segmented characters;

performing distance computations on the recognized characters using a Generalized Hidden Markov Model with ergodic topology to identify words based on character models;

performing non-keyword rejection using a classifier based on a combination of Gaussian Mixture Models, Hidden Markov Models and Support Vector Machines;

outputting the spotted keywords.

2. The method of claim 1, wherein the step of performing word segmentation on the image includes the sub-steps of:

performing pre-processing on the image to remove non-text elements;

performing line extraction on the pre-processed image to provide extracted text lines;

performing word segmentation on the extracted text lines.

3. The method of claim 2, wherein the sub-step of performing pre-processing on the image includes at least one operation selected from the group consisting of margin removal, skew correction, rule-line removal and noise removal.

4. The method of claim 2, wherein the sub-step of performing line extraction on the pre-processed image includes the sub-steps of:

applying a set of Fast Fourier Steerable filters to an adaptive local connectivity map of the pre-processed image using linear interpolation, the filters each being steered in different directions;

binarizing the filtered image to obtain a binarized line map;

post-processing the binarized line map to remove thin connected components;

vertically filling remaining connected components;

providing the extracted text lines.

5. The method of claim 2, wherein the sub-step of performing line extraction on the pre-processed image includes the sub-steps of:

applying a set of Fast Fourier Steerable filters to an adaptive local connectivity map of the pre-processed image using nearest-neighbor interpolation, the filters each being steered in different directions;

binarizing the filtered image to obtain a binarized line map;

post-processing the binarized line map to remove thin connected components;

providing the extracted text lines.

6. The method of claim 2, wherein the sub-step of performing line extraction on the pre-processed image further includes the sub-step of:

performing local skew correction inside a bounding box corresponding to each extracted text line.

7. The method of claim 2, wherein the step of performing word segmentation on the extracted text lines is performed with a scale-space algorithm using derivatives of fast anisotropic Gaussian filters implemented in the Fourier domain.

8. The method of claim 2, further comprising the steps of:

generating, prior to the step of performing word matching, word segmentation hypotheses for each of the extracted text lines using a Markov Chain;

performing a threshold selection on the word segmentation hypotheses using a segmentation threshold;

selecting the most likely word segmentation hypotheses based on the segmentation threshold.

9. The method of claim 2, further comprising the steps of:

generating, prior to the step of performing word matching, word segmentation hypotheses for each of the extracted text lines using a Hidden Markov Model;

performing a threshold selection on the word segmentation hypotheses using a segmentation threshold;

selecting the most likely word segmentation hypotheses based on the segmentation threshold.

10. The method of claim 8, wherein the segmentation threshold is set to a value that minimizes the error between the probabilities corresponding to perfectly segmented lines, over-segmented lines and under-segmented lines.

11. The method of claim 8, wherein the step of performing a threshold selection on the word segmentation hypotheses further uses automatically generated over and under segmentation training data.

12. The method of claim 11, wherein the over and under segmentation training data is obtained by performing the steps of:

providing a collection of training documents containing text line and word information;

computing estimates of average intra-word and inter-word distances;

generating over-segmented text lines by merging neighboring connected components closer than given first threshold percentage of the average intra-word distance;

generating over-segmented text lines by merging neighboring connected components closer than given second threshold percentage of the average inter-word distance.

15. The method of claim 1, wherein the step of performing word matching further comprises the sub-step of:

merging, prior to the sub-step of performing character recognition, broken characters in the segmented words.

16. The method of claim 15, wherein the sub-step of merging broken characters comprises the sub-steps of:

creating a sequence of connected components corresponding to characters or pieces of character in the segmented words;

defining a neighborhood relation based on the Euclidian distance and overlapping of the connected components;

creating a neighboring graph of the connected components based on the neighborhood relation;

partitioning the neighboring graph using an iterative using a best score based on the width and the overlapping of connected components;

merging broken characters based on the partitioned neighboring graph.

17. The method of claim 1, wherein the sub-step of performing character recognition on the segmented characters is based on a background skeletal graph.

18. The method of claim 17, wherein the sub-step of performing character segmentation on the segmented words includes the sub-steps of:

performing pre-processing on the segmented characters;

computing a background skeleton;

computing the skeletal graph from the background skeleton;

removing curves and short branches from the skeletal graph;

connecting each branch corresponding to an end-point of a downward branch that goes below a baseline of the image to a nearest point in the skeletal graph that is below the baseline of the image;

connecting each branch corresponding to an end-point of an upward branch that goes above the baseline of the image to a nearest point in the skeletal graph that is above the baseline of the image;

removing all remaining branches of the skeletal graph;

providing the segmented characters.

19. The method of claim 18, wherein the sub-step of performing character segmentation on the segmented words further includes the sub-step of:

detecting, prior to performing pre-processing on the segmented characters, under-segmented pairs of characters using a classifier based on a fuzzy interference system using average number of transition features.

20. The method of claim 18, wherein the sub-step of performing pre-processing on the segmented characters includes at least one operation selected from the group consisting of slant correction, binarization, vertical filling inside each connected components and removing isolated blocks.

21. The method of claim 18, wherein the sub-step performing pre-processing on the segmented characters is performed using an input perturbation and classifier combination algorithm.

13-14. -14. (canceled)

22. A method for improving word segmentation of a document, comprising the steps of:

obtaining extracted text lines from an image of the document;

generating word segmentation hypotheses for each of the extracted text lines using one of a Markov Chain and a Hidden Markov Model;

performing a threshold selection on the word segmentation hypotheses using a segmentation threshold;

selecting the most likely word segmentation hypotheses based on the segmentation threshold;

providing the segmented words.

24. The method of claim 22, wherein the step of performing a threshold selection on the word segmentation hypotheses further uses automatically generated over and under segmentation training data, wherein the over and under segmentation training data is obtained by performing the steps of:

providing a collection of training documents containing text line and word information;

computing estimates of average intra-word and inter-word distances;

generating over-segmented text lines by merging neighboring connected components closer than given first threshold percentage of the average intra-word distance;

generating over-segmented text lines by merging neighboring connected components closer than given second threshold percentage of the average inter-word distance.

23. (canceled)

25-54. -54. (canceled)